TO THE RANDOM WALK - Universitat de Barcelona - Home · TO THE RANDOM WALK Daniel Campos (Universitat Autònoma de Barcelona) 1. Introduction (Random search theory: applications and

THE OPTIMAL WALK

TO THE RANDOM WALK

Daniel Campos (Universitat Autònoma de Barcelona)

1. Introduction (Random search theory: applications and tools)

2. The optimal walk to the (Lévy) walk

3. The optimal walk to the (intermittent) walk

4. The optimal walk to the (myopic) walk

5. The optimal walk to the (mortal) walk

6. The optimal walk to the (systematic?) walk

Contents


A

B

The first-passage problem:

How do we measure search efficiency? i) Time distribution to the first passage: 𝑓(𝑡; 𝑥0)

ii) Mean time to the first passage: 𝑇 = 𝑡𝑓 𝑡; 𝑥0 𝑑𝑡∞

0

iii) First-passage probability up to time 𝑡𝑚: 𝑆 𝑡𝑚 = 𝑓 𝑡; 𝑥0 𝑑𝑡𝑡𝑚0


𝐿

A A

A

A

A

A

A

B B

B

B

B

B

Random search theory: 𝐴 + 𝐵 → 𝐴 We will describe the position of the 𝑖-th particle through a stochastic process 𝑋𝑖(𝑡).


𝐿

A A

A

A

A

A

A

B B

B

B

B

B

The target problem:


𝐿

A A

A

A

A

A

A

B B

B

B

B

B

The trapping problem:


𝐿


Movement ecology represents a new area of ecology which requires a detailed data processing of individual animal trajectories (obtained through telemmetry, GPS,…).


Search at the microscopic level


Human searches:


SAR applications

Experiments

Everyday experiences

A Wiener process 𝑊(𝑡) is defined as a stationary process whose increments 𝑊(𝑡2) −𝑊(𝑡1) follow a Gaussian distribution with zero mean and variance 𝑡2 − 𝑡1 .

If we assume that 𝑋 𝑡 = 𝑥0 + 2𝐷𝑊(𝑡) then: i) The probability density 𝑝(𝑥, 𝑡) follows a Gaussian distribution with 𝑋 = 𝑥0 and

𝑋2 = 2𝐷𝑡 + 𝑥02

ii) It becomes impossible to define a characteristic speed for A iii) The problem of infinite propagation signals emerge

…but the advantage is that we can describe 𝑋 𝑡 as a Gaussian (stable) process.

A

Types of motion (I): ‘Pure’ diffusion model


We define the position of the particle after 𝑛 jumps as: 𝑋𝑛 = 𝑍𝑖𝑛𝑖=1

…and the time it takes to perform these 𝑛 jumps as: 𝑇𝑛 = Θ𝑖𝑛𝑖=1

…where 𝑍𝑖 and Θ𝑖 each are i.i.d. random variables distributed, respectively, according to

ϕ 𝑥 : Jump-length probability distribution function (dispersal kernel) 𝜑 𝑡 : Waiting-time probability distribution function

(This is typically known as a Continuous-Time Random Walk –CTRW- and includes the Lévy Flight case as a particular case)

A

Types of motion (II): ‘Jump’ model


We use the same definition as before 𝑋𝑛 = 𝑍𝑖𝑛𝑖=1 𝑇𝑛 = Θ𝑖

𝑛𝑖=1

…where now 𝜑 𝑡 and ϕ 𝑥 are not independent, but coupled through a velocity distribution ℎ(𝑣) in the form

ϕ 𝑥 = 𝑑𝑡∞

0

𝜑(𝑡) 𝑑𝑡 𝛿 𝑥 − 𝑣𝑡 ℎ(𝑣)∞

−∞

(This is typically known as the “velocity version” of the CTRW, and includes the Lévy walk case, together with some other that ‘mimic’ the Ornstein-Uhlenbeck process in 𝑣)

A

Types of motion (III): ‘Velocity’ model

CTRW Velocity CTRW

𝑣1 𝑣2

𝑣3


𝑋(𝑡)

Direct resolution: We formally define the problem as 𝐿𝐹𝑃 𝑝(𝑥, 𝑡) = 0 with boundary condition 𝑝(𝛺, 𝑡) = 0, being 𝛺 the surface of the target, and computing the flux at 𝛺. For the

Wiener process, for example, 𝐿𝐹𝑃 =𝜕

𝜕𝑡− 𝐷

𝜕2

𝜕𝑥2.

Master equation approach: For the Bernoulli random walk (with probability ½ to each side, jump size 𝑎 and waiting time 𝜏) we have

𝑇 𝑥0 =1

2𝑇 𝑥0 + 𝑎 + 𝜏 +

1

2𝑇 𝑥0 − 𝑎 + 𝜏

𝑎2

2𝜏 𝜕2 𝑇 (𝑥0)

𝜕𝑥02

= −1

𝑇 (𝑥0) =𝑥0 𝐿 − 𝑥0 𝜏

𝑎2

Methods for finding 𝑓(𝑡) and/or 𝑇


Methods for finding 𝑓(𝑡) and/or 𝑇

The renewal approach We assume one target located at 𝑥 = 𝑥𝑡 and introduce 𝑘𝑛(𝑡; 𝑥0) as the rate at which the 𝑛-th passage occurs. Using a renewal assumption we have

𝑘 𝑡; 𝑥0 𝑑𝑡 = 𝑘1 𝑡; 𝑥0 𝑑𝑡 + 𝑘2 𝑡; 𝑥0 𝑑𝑡 + 𝑘3 𝑡; 𝑥0 𝑑𝑡 + ⋯ = = 𝑓 𝑡; 𝑥0 + 𝑓 𝑡; 𝑥0 ∗ 𝑓 𝑡; 𝑥𝑡 + 𝑓 𝑡; 𝑥0 ∗ 𝑓 𝑡; 𝑥𝑡 ∗ 𝑓 𝑡; 𝑥𝑡 +⋯ 𝑑𝑡

𝑘 𝑠; 𝑥0 = 𝑓 𝑠; 𝑥0 𝑓 𝑠; 𝑥𝑡𝑖

∞

𝑖=0

=𝑓(𝑠; 𝑥0)

1 − 𝑓(𝑠; 𝑥𝑡)

𝑓 𝑠; 𝑥0 =𝑘(𝑠; 𝑥0)

1 + 𝑘(𝑠; 𝑥𝑡)


𝑻𝟏

An essential advantage of this framework is that it allows a very general and intuitive understanding of the Mean First Passage Time (MFPT):

𝑇 = 𝑑𝑡 𝑡𝑓(𝑡; 𝑥0)∞

0

= lim𝑠→0

𝑑𝑓(𝑠; 𝑥0)

𝑑𝑠= lim

𝑠→0

𝑘∗ 𝑠; 𝑥𝑡𝑘(∞)

−𝑘∗(𝑠; 𝑥0)

𝑘 ∞+

1

𝑘(∞)

𝑻𝟐



WHAT ARE LEVY FLIGHTS AND LEVY WALKS?

The Lévy Flight fits our ‘jump’ model scheme with ϕ 𝑥 a jump length distribution which decays according to lim

𝑡→∞ϕ 𝑥 ~𝑥−𝜇, with 1 < 𝜇 < 3

The Lévy Walk fits our ‘velocity’ model scheme, with 𝑣 fixed and 𝜑 𝑡 a flight time distribution which decays according to lim

𝑡→∞𝜑 𝑡 ~𝑡−𝜇, with 1 < 𝜇 < 3.

Note that this implies that 𝑥𝑞 ≡ 𝑑𝑥ϕ 𝑥 𝑥𝑞∞

−∞ and 𝑡𝑞 ≡

𝑑𝑡𝜑(𝑡)𝑡𝑞∞

0, respectively, diverge for 𝑞 − 𝜇 ≥ −1

In the Lévy Flight case these divergences extend to the overall behavior of the particle, so 𝑋2 also diverges. In contrast, for the Lévy Walk case, thanks to the coupling between flight durations and lengths through 𝑣:

𝑋2 ~ 𝑡2 , 1 < 𝜇 < 2

𝑡4−𝜇 , 2 < 𝜇 < 3



THE LÉVY FLIGHT OPTIMAL SOLUTION

Define the search efficiency 1

𝑙 𝑁 , where 𝑙 is the mean flight distance between

targets and 𝑁 the mean number of flights to cover the distance between targets Given ϕ 𝑥 ~𝑥−𝜇 and a mean path between targets of 𝛽,

𝑙 ≈ 𝑑𝑥 𝑥1−𝜇𝛽

0+ 𝛽 𝑑𝑥 𝑥−𝜇∞

𝛽

𝑑𝑥 𝑥−𝜇∞

0

and the mean number of flights satisfies 𝑁~𝛽(𝜇−1)/2 if the target is close enough. All this leads to a search efficiency optimization for 𝜇 = 2. INTUITIVE MEANING Optimal ballistic approach Optimal reorientation (‘correction’)

0 𝐿 0 𝐿

(Viswanathan et. al. Nature 401, 911 (1999)


THE LEVY FLIGHT FORAGING HYPOTHESIS “Given Lévy Flight optimality, evolution should have favoured sensorymotor mechanisms that facilitate the emergence of motion patterns similar to the Lévy case in search situations of poor information.”

ARE LÉVY FLIGHTS SO SPECIAL?



Practical example 1: DNA facilitated target location (1D sliding + jumping)


Practical example 2: Saltatory search strategies


O. Bénichou et. al. Rev. Mod. Phys. 83, 81 (2011)


Static mode (1D)

For a domain of size b:


Diffusive mode (1D) 𝑏𝐷2 < 𝑎3𝑉2 𝑏𝐷2 > 𝑎3𝑉2


Diffusive mode (1D) O. Bénichou et. al. Phys. Rev. Lett. 94, 198101 (2005)

𝜏1 ≫ 𝐷/𝑉2 𝜏1 ≪ 𝐷/𝑉2



For a constant probability 𝛼 of detection, 𝑘 𝑡; 𝑥0 → 𝛼𝑘(𝑡; 𝑥0) For a speed-dependent probability 𝛼 = 𝛼(𝑣): In general, 𝛼 = 𝛼(𝑥, 𝑣, 𝑡):

𝑘 𝑡; 𝑥0 𝑑𝑡 = 𝑑𝑣∞

0

𝑑𝑥0

0−𝑣𝑑𝑡

𝛼(𝑣)𝑝 𝑥, 𝑣, 𝑡; 𝑥0 + 𝑑𝑣0

−∞

𝑑𝑥𝑣𝑑𝑡

0

𝛼(𝑣)𝑝 𝑥, 𝑣, 𝑡; 𝑥0 ≈

≈ 𝑑𝑣∞

0

𝑣𝛼(𝑣)𝑝 0, 𝑣, 𝑡; 𝑥0 𝑑𝑡 + 𝑑𝑣0

−∞

𝑣𝛼(𝑣)𝑝 0, 𝑣, 𝑡; 𝑥0 𝑑𝑡

𝑘 𝑡; 𝑥0 𝑑𝑡 ≈ 𝑑𝑣∞

0

𝑣𝛼 𝑥, 𝑣, 𝑡 𝑝 0, 𝑣, 𝑡; 𝑥0 𝑑𝑡 − 𝑑𝑣0

−∞

𝑣𝛼(𝑥, 𝑣, 𝑡)𝑝 0, 𝑣, 𝑡; 𝑥0 𝑑𝑡


𝑻𝟏

𝑻𝟐

Case 1D ‘velocity’ model with 𝑣 fixed and 𝜑 𝑡 = 𝜆𝑒−𝜆𝑡 or 𝜑 𝑠 = 𝑒−𝑎 𝑘 𝜇, 𝛼 = 𝑒−𝛾𝑣

𝑇 =2𝑥0 𝐿−𝑥0 𝜆

𝑣2+

𝐿

𝑣𝛼(𝑣)


𝑻𝟏

𝑻𝟐

Case 2D or higher ‘velocity’ model with 𝑣 fixed and 𝜑 𝑡 = 𝜆𝑒−𝜆𝑡, 𝛼 = 𝑒−𝛾𝑣

𝑇 =𝐿2

𝑣2𝑔𝑑 𝑥0 +

𝐿𝑑

𝐴𝑣𝛼(𝑣)=

𝐿2

𝑣2𝑔𝑑 𝑥0 +

1

𝜌

1

𝐴𝑣𝛼(𝑣)


Case 2D with 𝑣 fixed and 𝜑 𝑡 = 𝜆𝑒−𝜆𝑡 or 𝜑 𝑠 = 𝑒−𝑎 𝑘 𝜇, 𝛼 = 𝑒−𝛾𝑣


For a constant mortality rate 𝜔: Here 𝑆(∞) is the most relevant parameter: For the ‘pure’ diffusion model:

For the ‘velocity’ model with 𝑣 fixed and 𝜑 𝑡 = 𝜆𝑒−𝜆𝑡:

𝑘 𝑡; 𝑥0 → 𝑒−𝜔𝑡𝑘(𝑡; 𝑥0) → 𝑘 𝑠; 𝑥0 → 𝑘 𝑠 + 𝜔; 𝑥0 → 𝑓 𝑠; 𝑥0 =𝑘(𝑠+𝜔;𝑥0)

1+𝑘(𝑠+𝜔;𝑥𝑡)

𝑆 ∞ = 𝑑𝑡 𝑓(𝑡; 𝑥0)∞

0

= lim𝑠→0

𝑑𝑡 𝑒−𝑠𝑡𝑓(𝑡; 𝑥0)∞

0

= lim𝑠→0

𝑓(𝑠; 𝑥0)

𝑆 ∞ = 1 −𝜔(𝜔 + 𝜆) 𝑒− 𝜔 𝜔+𝜆 𝑥0/𝑣 + 𝑒− 𝜔 𝜔+𝜆 𝐿−𝑥0 /𝑣

𝜔 1 − 𝑒− 𝜔 𝜔+𝜆 𝐿/𝑣 + 𝜔(𝜔 + 𝜆) 1 + 𝑒− 𝜔 𝜔+𝜆 𝐿/𝑣

𝑆 ∞ = 1 −𝜔𝜆 𝑒− 𝜔𝜆 𝑥0/𝑣 + 𝑒− 𝜔𝜆 𝐿−𝑥0 /𝑣

1 + 𝑒− 𝜔𝜆 𝐿/𝑣


Implications for the Lévy flight paradigm:



NON-PERFECT DETECTION: SACCADE-FIXATION MECHANISM

𝑐 = 1 𝑐 = 0.2



Optimal search theory:

“Random” search


Optimal effort allocation vs optimal path planning

Parallel Sweep (PS)

Archimedean spiral (AS)


Errors in pattern sizing


Errors in prior information


Example: Ellberg’s paradox

30 balls Red: 10 Yellow+Black: 20

Does it make any sense to foresee human errors? According to psychologycal research during the last 50 years, yes


Collaborators: Vicenç Méndez

Universitat Autònoma de Barcelona

Frederic Bartumeus Centre d’Estudis Avançats de Blanes, Girona

Xavier Espadaler

Jordi Catalán John Palmer

Centre de Recerca Ecològica i Aplicacions Forestals, UAB

Enrique Abad Santos Bravo Yuste

Universidad de Extremadura

Will Ryu University of Toronto

José Soares de Andrade Jr

Universidade Federal do Ceará

Katja Lindenberg University of California San Diego

TO THE RANDOM WALK - Universitat de Barcelona - Home · TO THE RANDOM WALK Daniel Campos (Universitat Autònoma de Barcelona) 1. Introduction (Random search theory: applications and

Documents